Diverse Misinformation Impacts of Human Biases on Detection of Deepfakes on Networks Juniper Lovato12 Jonathan St-Onge1 Randall Harp13 Gabriela Salazar Lopez1 Sean P.

2025-05-03 0 0 3.19MB 19 页 10玖币

侵权投诉

Diverse Misinformation: Impacts of Human Biases on

Detection of Deepfakes on Networks

Juniper Lovato1,2,*, Jonathan St-Onge1, Randall Harp1,3, Gabriela Salazar Lopez1, Sean P.

Rogers1, Ijaz Ul Haq2, Laurent H´

ebert-Dufresne1,2, and Jeremiah Onaolapo1,2

1Vermont Complex Systems Center, University of Vermont, Burlington, 05405, U.S.A.

2Department of Computer Science, University of Vermont, Burlington, 05405, U.S.A.

3Department of Philosophy, University of Vermont, Burlington, 05405, U.S.A.

*juniper.lovato@uvm.edu

ABSTRACT

Social media platforms often assume that users can self-correct against misinformation. However, social media users are not

equally susceptible to all misinformation as their biases inﬂuence what types of misinformation might thrive and who might be

at risk. We call “diverse misinformation” the complex relationships between human biases and demographics represented in

misinformation. To investigate how users’ biases impact their susceptibility and their ability to correct each other, we analyze

classiﬁcation of deepfakes as a type of diverse misinformation. We chose deepfakes as a case study for three reasons: 1) their

classiﬁcation as misinformation is more objective; 2) we can control the demographics of the personas presented; 3) deepfakes

are a real-world concern with associated harms that must be better understood. Our paper presents an observational survey

(N=2,016) where participants are exposed to videos and asked questions about their attributes, not knowing some might be

deepfakes. Our analysis investigates the extent to which different users are duped and which perceived demographics of

deepfake personas tend to mislead. We ﬁnd that accuracy varies by demographics, and participants are generally better at

classifying videos that match them. We extrapolate from these results to understand the potential population-level impacts of

these biases using a mathematical model of the interplay between diverse misinformation and crowd correction. Our model

suggests that diverse contacts might provide “herd correction” where friends can protect each other. Altogether, human

biases and the attributes of misinformation matter greatly, but having a diverse social group may help reduce susceptibility to

misinformation.

1 Introduction

There is a growing body of scholarly work focused on distributed harm in online social networks. From leaky data,

and group

security and privacy

to hate speech,

misinformation

and detection of computer-generated content.

Social media users are

not all equally susceptible to these harmful forms of content. Our level of vulnerability depends on our own biases. We deﬁne

“diverse misinformation” as the complex relationships between human biases and demographics represented in misinformation.

This paper explores deepfakes as a case study of misinformation to investigate how U.S. social media users’ biases inﬂuence

their susceptibility to misinformation and their ability to correct each other. We choose deepfakes as a critical example of the

possible impacts of diverse misinformation for three reasons: 1) their status of being misinformation is binary; they either

are a deepfake or not; 2) the perceived demographic attributes of the persona presented in the videos can be characterized by

participants; 3) deepfakes are a current real-world concern with associated negative impacts that need to be better understood.

Together, this allows us to use deepfakes as a critical case study of diverse misinformation to understand the role individual

biases play in disseminating misinformation at scale on social networks and in shaping a population’s ability to self-correct.

We present an empirical survey (N=2,016 using a Qualtrics survey panel

) observing what attributes correspond to U.S.-

based participants’ ability to detect deepfake videos. Survey participants entered the study under the pretense that they would

judge the communication styles of video clips. Our observational study is careful not to prime participants at the time of their

viewing video clips so we could gauge their ability to view and judge deepfakes when they were not expecting them (not

explicitly knowing if a video is fake or not is meant to emulate what they would experience in an online social media platform).

Our survey also investigates the relationship between human participants’ demographics and their perception of the video

person(a)’s features and, ultimately, how this relationship may impact the participant’s ability to detect deepfake content.

Our objective is to evaluate the relationship between classiﬁcation accuracy and the demographic features of deepfake

videos and survey participants. Further analysis of other surveyed attributes will be explored in future work. We also recognize

that data used to train models that create deepfakes may introduce algorithmic biases in the quality of the videos themselves,

which could introduce additional biases in the participant’s ability to guess if the video is a deepfake or not.

arXiv:2210.10026v3 [cs.SI] 13 Jan 2024

The Facebook Deepfake Detection Challenge dataset that was used to create the videos we use in our survey was created

to be balanced in diversity in several axes (gender, skin-tone, age). We suspect that if there are algorithmic-level biases in

the model used resulting in better deepfakes for personas of speciﬁc demographics, we would expect to see poorer accuracy

across the board for all viewer types when classifying these videos. We do see that viewer groups’ accuracy differs based on

different deepfake video groups. However, our focus is on the perception of survey participants towards deepfakes’ identity

and demographics to capture viewer bias based on their perception rather than the model’s bias and classiﬁcation of the video

persona’s racial, age, and gender identity. Our goal is to focus on viewers and capture what a viewer would experience in the

wild (on a social media platform), where a user would be guessing the identity features of the deepfake and then interrogating if

the video was real or not with little to no priming.

Figure 1. Illustration of the problem considered in this work. Populations are made of individuals with diverse demographic

features (e.g., age, gender, race; here represented by colors), and misinformation is likewise made of different elements based

on the topics they represent (here shown as pathogens). Through their biases, certain individuals are more susceptible to certain

kinds of misinformation. The cartoon represents a situation where misinformation is more successful when it matches an

individual’s demographic. Red pathogens spread more readily around red users with red neighbors, thereby creating a

misinformed echo chamber whose members can not correct each other. In reality, the nature of these biases is still unclear, and

so are their impacts on online social networks and on the so-called “self-correcting crowd”.

This paper adopts a multidisciplinary approach to answer these questions and understand their possible impacts. First, we

use a survey analysis to explore individual biases related to deepfake detection. There is abundant research suggesting the

demographics of observers and observed parties inﬂuence the observer’s judgment and sometimes actions toward the observed

party.

7–11

In an effort to avoid assumptions about any demographic group, we chose four speciﬁc biases to analyze vis-à-vis

deepfakes: (Question 1) Priming bias: How much does classiﬁcation accuracy depend on participants being primed about the

potential of a video being fake? Our participants are not primed on the meaning of deepfakes and are not told to be explicitly

looking for them prior to beginning the survey. Importantly, we do not explicitly vary the priming of our participants but we

compare their accuracy to a previous study with a similar design but primed participants

. Participants are debriefed after the

completion of the survey questions and then asked to guess the deepfake status of the videos they watched. More information

about our survey methodology and why the study was formulated as a deceptive survey can be seen in section 4.3. (Question 2)

Prior knowledge: Does accuracy depend on how often the viewer uses social media and whether they have previously heard of

deepfakes? Here, we ask participants to evaluate their own knowledge and use their personal assessment to answer this research

question. (Question 3) Homophily bias: Are humans better classiﬁers of video content if the perceived demographic of the

video persona matches their own identity? (Question 4) Heterophily bias: Inversely, are humans more accurate if the perceived

demographic of the video persona does not match their own? We then use results from the survey to develop an idealized

mathematical model to theoretically explore population-level dynamics of diverse misinformation on online social networks.

Altogether, this allows us to hypothesize the mechanisms and possible impacts of diverse misinformation, as illustrated in

Fig. 1.

Our paper is structured as follows. We outline the harms and ethical concerns of diverse misinformation and deepfakes

in Section 1. We explore the possible effects through which demographics impact susceptibility to diverse misinformation

through our observational study in Section 2. We then investigate the network-level dynamics of diverse misinformation using a

mathematical model in Section 2.1. We discuss our ﬁndings and their implications in Section 3. Our full survey methodology

2/19

can be seen in Section 4.

Subsequently, these biases impact how social ties are formed and, ultimately, the shape of the social network. For example,

in online social networks, homophily often manifests through triadic closures

where friends in social networks tend to

form new connections that close triangles or triads. Understanding individuals’ and groups’ biases will help understand the

network’s structure and dynamics and how information and misinformation spread on the network depending on its level of

diversity. For example, depending on the biases and the node-speciﬁc diversity of the connections it forms, one may have a

system that may be more or less susceptible to widespread dissemination as it would in a Mixed Membership Stochastic Block

Model (MMSBM)

. A Mixed Membership Stochastic Block Model is a Bayesian community detection method that segments

communities into blocks but allows community members to mix with other communities. Assumptions in an MMSBM include

a list of probabilities that determine the likelihood of communities interacting. We explore these topics in more detail in Section

2.1.

Previous work has demonstrated that homophily bias towards content aligned with one’s political afﬁliation can impact

one’s ability to detect misinformation.

14,15

Traberg et al. show that political afﬁliation can impact a person’s ability to detect

misinformation about political content.

They found that viewers misclassiﬁed misinformation as being true more often when

the source of information aligned with their political afﬁliation. Political homophily bias, in this case, made them feel as though

the source was more credible than it was.

In this paper, we investigate the accuracy of deepfake detection based on multiple homophily biases in age, gender, and race.

We also explore other bias types, such as heterophily bias, priming, and prior knowledge bias impacting deepfake detection.

Misinformation is information that imitates real information but does not reﬂect the genuine truth.

Misinformation

has become a widespread societal issue that has drawn considerable recent attention. It circulates physically and virtually

on social media sites

and interacts with socio-semantic assortativity. In contrast, assortative social clusters will also tend

to be semantically homogeneous.

For instance, misinformation promoting political ideology might spread more easily in

social clusters based on shared demographics, further exacerbating political polarization and potentially inﬂuencing electoral

outcomes

. This has sparked concerns about the weaponization of manipulated videos for malicious ends, especially in the

political realm

. Those with higher political interests are more likely to share deepfakes inadvertently, and those with lower

cognitive ability are also more likely to share deepfakes inadvertently. The relationship between political interest and deepfakes

sharing is moderated by network size20.

Motivations vary broadly to explain why people disseminate misinformation, which we refer to as disinformation when

speciﬁcally intended to deceive. Motivations include 1) purposefully trying to deceive people by seeding distrust in information,

2) believing the information to be accurate and spreading it mistakenly, and 3) spreading misinformation for monetary gain. In

this paper, we will primarily focus on deepfakes as misinformation meaning the potential of a deepfake viewer getting duped

and sharing a deepfake video. Disinformation is spreading misinformation with the intent to deceive. In this paper, we do not

assume that all deepfakes are disinformation since we do not consider the intent of the creator. A deepfake could be made to

entertain or showcase technology. We instead focus on deepfakes as misinformation meaning the potential of a deepfake viewer

getting duped and sharing a deepfake video, regardless of intent.

There are many contexts where online misinformation is of concern. Examples include misinformation around political

elections and announcements (political harms)

; such deepfake videos can, in theory, alter political ﬁgures to say just about

anything, raising a series of political and civic concerns

; misinformation on vaccinations during global pandemics (health-

related harms)

22,23

; false speculation to disrupt economies or speculative markets

; distrust in news media and journalism

(harms to news media)

4,25

. People are more likely to feel uncertain than to be misled by deepfakes, but this resulting uncertainty,

in turn, reduces trust in news on social media

; false information in critical informational periods such as humanitarian or

environmental crises

; and propagation of hate speech online

which spreads harmful false content and stereotypes about

groups (harms related to hate speech).

Correction of misinformation: There are currently many ways to try to detect and mitigate the harms of misinformation

online.

On one end of the spectrum are automated detection techniques that focus on the classiﬁcation of content or on

observing anomaly detection in the network structure context of the information or propagation patterns.

29,30

Conversely,

crowd-sourced correction of misinformation leverages other users to reach a consensus or simply estimate the veracity of the

content.

31–33

We will look at the latter form of correction in an online social network to investigate the role group correction

plays in slowing the dissemination of diverse misinformation at scale.

Connection with deepfakes: The potential harms of misinformation can be ampliﬁed by computer-generated videos used

to give fake authority to the information. Imagine, for instance, harmful messages about an epidemic conveyed through the

computer-generated persona of a public health ofﬁcial. Unfortunately, deepfake detection remains a challenging problem, and

the state-of-the-art techniques currently involve human judgment.5

Deepfakes are artiﬁcial images or videos in which the persona in the video is generated synthetically. Deepfakes can be

seen as false depictions of a person(a) that mimics a person(a) but does not reﬂect the truth. Deepfakes should not be confused

3/19

with augmented or distorted video content, such as using color ﬁlters or digitally-added stickers in a video. Creating a deepfake

can involve complex methods such as training artiﬁcial neural networks known as generative adversarial networks (GANs) on

existing media34 or simpler techniques such as face mapping.

Deepfakes are deceptive tools that have gained attention in recent media for their use of celebrity images and their ability to

spread misinformation across online social media platforms.35

Early deepfakes were easily detectable with the naked eye due to their uncanny visual attributes and movement.

However,

research and technological developments have improved deepfakes, making them more challenging to detect.

There are

currently several automated deepfake detection methods.

37–41

However, they are computationally expensive to deploy at scale.

As deepfakes become ubiquitous, it will be necessary for the general audience to identify deepfakes independently during gaps

between the development of automated techniques or in environments that are not always monitored by automated detection (or

are ofﬂine). It will also be important to allow human-aided and human-informed deepfake detection in concert with automated

detection techniques.

Several issues currently hinder automated methods: 1) they are computationally expensive; 2) there may be bias in deepfake

detection software and training data—credibility assessments, particularly in video content, have been shown to be biased;

3) As we have seen with many cybersecurity issues, there is a “cat-and-mouse” evolution that will leave gaps in detection

methodology.43

Humans may be able to help ﬁll these detection gaps. However, we wonder to what extent human biases impact the efﬁcacy

of detecting diverse misinformation. If human-aided deepfake detection becomes a reliable strategy, we need to understand the

biases that come with it and what they look like on a large scale and on a network structure. We also posit that insights into

human credibility assessments of deepfakes could help develop more lightweight and less computationally expensive automated

techniques.

As deepfakes improve in quality, the harms of deepfake videos are coming to light.

Deepfakes raise several ethical

considerations: 1) the evidentiary power of video content in legal frameworks;

4,45,46

2) consent and attribution of the

individual(s) depicted in deepfake videos;

3) bias in deepfake detection software and training data;

4) degradation of our

epistemic environment, i.e., there is a large-scale disagreement between what community members believe to be real or fake,

including an increase in misinformation and distrust;4,25 and 5) possible intrinsic wrongs of deepfakes.48

It is important to understand who gets duped by these videos and how this impacts people’s interaction with any video

content. The gap between convincing deepfakes and reliable detection methods could pose harm to democracy, national security,

privacy, and legal frameworks.

Consequently, additional regulatory and legal frameworks

will need to be adopted to protect

citizens from harms associated with deepfakes and uphold the evidentiary power of visual content. False light is a recognized

invasion of privacy tort that acknowledges the harms that come when a person has untrue or misleading claims made about

them. We suspect that future legal protections against deepfakes might well be grounded in such torts, though establishing these

legal protections is not trivial.45,50

The ethical implications of deepfake videos can be separated into two main categories: the impacts on our epistemic

environment and people’s moral relationships and obligations with others and themselves. Consider the epistemic environment,

which includes our capacity to take certain representations of the world as true and our taking beliefs and inferences to be

appropriately justiﬁed. Audio and video are particularly robust and evocative representations of the world. They have long been

viewed as possessing more testimonial authority (in the broader, philosophical sense of the phrase) than other representations

of the world. This is true in criminal and civil contexts in the United States, where the admissibility of video recordings as

evidence in federal trials is speciﬁcally singled out in Article X of the Federal Rules of Evidence

(State courts have their own

rules of evidence, but most states similarly have explicit rules that govern the admissibility of video recordings as evidence).

The wide adoption of deepfake technology would strain these rules of evidence; for example, the federal rules of evidence

reference examples of handwriting authentication, telephone conversation authentication, and voice authentication but do not

explicitly mention video authentication. Furthermore, laws are notorious for lagging behind technological advances,

which

can further complicate and limit how judges and juries can approach the existence of a deepfake video as part of a criminal or

civil case.

Our paper asks four primary research questions regarding how human biases impact deepfake detection. (Q1) Priming:

How important is it for an observer to know that a video might be fake? (Q2) Prior knowledge: How important is it for an

observer to know about deepfakes, and how does social media usage affect accuracy? (Q3-Q4) Homophily and heterophily

biases: Are participants more accurate at classifying videos whose persona they perceive to match (homophily) or mismatch

(heterophily) their own demographic attributes in age, gender, and race?

To address our four research questions, we designed an IRB-approved survey (N=2,016) using video clips from the Deepfake

Detection Challenge (DFDC) Preview Dataset.

53,54

Our survey participants entered the study under the pretense that they

would judge the communication styles of video clips (they were not explicitly looking for deepfake videos in order to emulate

the uncertainty they would experience in an online social network). After the consent process, survey participants were asked to

4/19

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

DiverseMisinformation:ImpactsofHumanBiasesonDetectionofDeepfakesonNetworksJuniperLovato1,2,*,JonathanSt-Onge1,RandallHarp1,3,GabrielaSalazarLopez1,SeanP.Rogers1,IjazUlHaq2,LaurentH´ebert-Dufresne1,2,andJeremiahOnaolapo1,21VermontComplexSystemsCenter,UniversityofVermont,Burlington,05405,U.S.A.2Depart...

展开>> 收起<<

Diverse Misinformation Impacts of Human Biases on Detection of Deepfakes on Networks Juniper Lovato12 Jonathan St-Onge1 Randall Harp13 Gabriela Salazar Lopez1 Sean P..pdf

共19页,预览4页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

Diverse Misinformation Impacts of Human Biases on Detection of Deepfakes on Networks Juniper Lovato12 Jonathan St-Onge1 Randall Harp13 Gabriela Salazar Lopez1 Sean P.

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: